IR after optimization increases code instructions in loop

[r00t0s]

The code before optimization is more shorter than after optimization, when I use -O1 code looks correct, but when I use -O2 -O3 looks so bad, for example GCC have similar problem but only when used -O3

Example: -O1
https://godbolt.org/z/9nq3hd8ve

After use -O2 or -O3
https://godbolt.org/z/xYeKzveze

All versions of clang have the same problem

[antoniofrighetto]

When disabling vectorizers, it looks like LoopUnrolling still makes the code worse: https://godbolt.org/z/35qM4zs6j.

[nikic]

I don't understand the problem. The code gets vectorized/unrolled -- isn't that expected? You asked to optimize for speed, not size (-Os).

[antoniofrighetto]

Yeah, probably a 4x-unrolling here has generally faster execution times than the non-unrolled version, though I was expecting O2 to optimize for speed whilst still not increasing the code size too much (size increase is probably fine here though).

[antoniofrighetto]

Yeah, probably a 4x-unrolling here has generally faster execution times than the non-unrolled version, though I was expecting O2 to optimize for speed whilst still not increasing the code size too much (size increase is probably fine here though).

Or, arguably, one might have expected the output above with O3 :)

[fhahn]

With O2/O3 we vectorize the loop with a vector loop to process 16 elements and another vectorized version to process 4 elements.

The main difference to GCC is that I think GCC doesn't enable vectorization at -O2. The vectorized code will be substantially faster.

[nikic]

I believe everything here is working as intended, so closing. Optimize for code size if you want to optimize for code size.